Precision 'unsigned' on Apple M1 leads to inaccurate reading of kernel arguments

[Steve73]

What is wrong?

When you use precision 'unsigned' on the Apple M1 GPU (with Safari and Chrome), kernel arguments are read inaccurately.

Where does it happen?

const gpu = new GPU({ mode: 'gpu' });

const createTexture1 = gpu.createKernel(function(r) {
  return r;
}, { output: [1], precision: 'unsigned'});

var t1 = createTexture1(2);
console.log(t1);

const createTexture2 = gpu.createKernel(function() {
  return 2;
}, { output: [1], precision: 'unsigned'});

var t2 = createTexture2();
console.log(t2);

This leads to this output:

Float32Array [2.015624761581421]
Float32Array [2]

Both lines should show the same value (2).

How do we replicate the issue?

See code example above.

How important is this (1-5)?

Not sure to what extent such inaccuracies are tolerated.

Expected behavior (i.e. solution)

Both output lines should show the same value (2).

[fuzzie360]

Fixed in 2.19.0. You were exactly right that unsigned precision was inaccurate on M1: the float→RGBA8 codec used log2/pow/exp2, which are slightly approximate on Apple GPUs — enough to corrupt the packed bytes (powers of two came back as e.g. 2.015625; everything else lost its low bits). The codec now uses exact arithmetic; a 198-value round-trip harness has zero failures on Apple silicon, and the suite there went from 157 failures to 12. npm publish pending; until then: npm install gpujs/gpu.js#2.19.0.